Confidential Mode
From its catalog listing
Hermes Confidential Mode locks a Hermes profile to one tier: attested (a Tinfoil enclave, or NEAR AI Cloud's gateway and models, verified and pinned to a build you approve) or on-host-only (loopback and hosts you list). While locked, the plugin refuses off-tier main model calls before they are sent; refuses to lock while a side call, fallback, memory, MCP URL or delegation route would leave the tier; on the attested tier encrypts request bodies to the verified enclave or model key and pins every inference connection to the TLS key of the attestation the plugin accepted (the attestation fetch itself is not pinned, by design), with headers, sizes and NEAR images still visible to the vendor; and asks you the first time each tool it does not class as on-host (by where the tool's data goes) runs in each conversation during a lock, blocking the call when nobody can answer. The plugin logs each model call's host and size, never its text, and renders a receipt whose attached enclave evidence replays offline; the rest of the receipt is the plugin's own record. The attested tier needs TINFOIL_API_KEY or NEAR_AI_API_KEY. Shell and code execution are not inspected for network access, and a side call that strays mid-turn is recorded, not stopped.
Good to know before installing
the plugin contacts the vendor (atc.tinfoil.sh, the Tinfoil router and the *.tinfoil.sh collateral proxies, or cloud-api.near.ai), Intel Trust Authority (portal and portal.eu), Sigstore's TUF CDN, and GitHub (raw.githubusercontent.com, api.github.com) for NEAR compose files, also in the background after a verification stages a NEAR entry; the plugin sends no telemetry. The plugin writes under <HERMES_HOME>/plugin-data/ (including the ledger, evidence blobs, pins.yaml, allowlist.yaml, state and SDK caches), Hermes' tool-discovery cache on a terminal lock, and, on a revoke only, config.yaml, removing its own plugin_rule:confidential-mode:tool entries from command_allowlist and no other entry, through Hermes' save_config, which also normalises legacy model and max_turns keys as any Hermes config save does; the plugin reads config.yaml and script files a locked tool call runs. The plugin reads TINFOIL_API_KEY and NEAR_AI_API_KEY through Hermes' secret scope, checks only whether OPENROUTER_API_KEY and NOUS_API_KEY are set (to report Hermes' side-call fall-through), reads TERMINAL_ENV, TERMINAL_CWD, HERMES_HOME, CONFIDENTIAL_MODE_PROGRESS and CONFIDENTIAL_MODE_NO_FETCH, sets XDG_CONFIDENTIAL_MODE_TOOL_PROCESS, stores no credential and runs no shell command or subprocess. The plugin starts daemon threads of five kinds: an attestation, a pins.yaml match record, a NEAR compose-file fetch, the own-host-name DNS lookup, and, while locked, one long-running re-verification thread per attested provider. During each Tinfoil attestation and each hermes confidential replay, the plugin wraps the Tinfoil SDK's fetch functions and requests.Session.request process-wide, inert for every other thread, and puts the original functions back afterwards; before its first verification it also re-points the Tinfoil SDK's cache-directory constants and sigstore-python's TUF directory function at its plugin-data directory for the rest of the process. The plugin adds a locked-mode note to each locked turn's context and may append one notice line to a reply. hermes plugins validate counts 31 private core import sites, all reads (docs/DISCLOSURE.md).
From its catalog listing.
How to install
- Install it. Hermes asks for any API keys it needs, then asks whether to turn it on.
hermes plugins install confidential-mode - Said no? Turn it on later:
hermes plugins enable confidential-mode - Restart Hermes so it loads the plugin.
Some plugins also need a sign-in or settings: see Full setup & settings. Memory plugins ask whether to become your memory provider instead. Steps follow the official plugins guide (opens in a new tab).
- Category
- Memory & notes · AI models & providers
- Maintained by
- ainsleys (Community)
- Needs
- No API keys declared
- Adds hooks
pre_tool_callpost_approval_responsepre_llm_callpre_api_requestpre_auxiliary_callpost_auxiliary_callon_session_starton_session_endpost_compactionpre_gateway_dispatchtransform_llm_output- Needs Hermes
- 0.22.7 or newer
- Version
- 0.1.0
- In the catalog
- 2026-11-10
Known issues (from its catalog listing)
- Auxiliary calls (compression, titles, vision and other side jobs) bypass the llm_execution middleware, so the plugin refuses to lock while one is routed off-tier and records a breach if one strays; it cannot stop the call.
- Hermes' CLI approve directive grants on any answer except 'deny'; the plugin's tool_execution middleware runs a tool only on an exact once, session or always.
- terminal, execute_code, process_manage and MCP servers launched as a command run on this host while locked and can reach the network; the plugin scans shell and code text only for its own commands, not for network access.
- Locked mode requires Hermes' local terminal backend: the lock is refused on any other backend (Docker included) and while Hermes holds a cached terminal environment on one.
- Locked mode does not apply in safe-worker mode, where Hermes loads no plugins.
- Every plugin in the same process shares the trust boundary; under plugins.isolation: host the attested provider is invisible and the lock is refused.
- Tinfoil attests its router, not the model backend behind it, and its report carries no client nonce; the plugin's own timestamps are the host clock.
- The plugin reads keys only through agent.secret_scope.get_secret, so a key held only in Hermes' credential pool (hermes auth) reads as not set and the lock refuses until the variable is set.
- AMD and Intel routers have different enclave fingerprints, so a restart can ask for a new pin; set confidential.tinfoil.enclave to one router to stop this.
- An 'always' tool grant lives in Hermes' command_allowlist and outlives the lock until /confidential revoke <tool>.
- After a runtime switch to a hosted provider, Hermes' start-up model-list request goes out with your API key (no conversation text) before any plugin seam runs; the plugin then refuses the turn.
- If Hermes runs on someone else's machine, that operator sees every prompt; locked mode protects the road out of the Hermes process, not the process.
- hermes confidential replay uses the evidence's own trust root and capture time unless you pass --trusted-root and --as-of, identifies the repository by name only unless you pass --expected-repo-id and --expected-owner-id, and the first pin is trust on first use.
- There is no mixed 'attested or on-host' tier yet, so side jobs cannot run on an on-host model while the main model is attested.
- The attested tier admits only models served inside Tinfoil or NEAR enclaves (open-weight families such as Qwen, DeepSeek, Llama and gpt-oss); no Claude or GPT model is in tier, and because fallbacks are refused while locked, a vendor outage ends the turn.
- While locked, tool calls not classed as on-host are blocked rather than asked under yolo, approvals.mode off, cron, one-shot (-q, -z) and unattended runs, since nobody can answer; 46 of Hermes' 73 built-in tools ask (docs/TOOL_CLASSES.md), the kanban writes and text_to_speech among them.
- NEAR AI Cloud: streaming is refused; images reach NEAR's gateway in plaintext unless confidential.near.image_parts: refuse; the gateway reads model id, parameters, headers, sizes and timing, can read a reply by swapping the client key header (caught only afterwards), and instances share one key (docs/LIMITATIONS.md section 14).
- An agent with terminal access on the same host can reach the approval prompt for a pin or a locked-allowlist tool by some routes; the ledger records every approval with its surface (docs/LIMITATIONS.md section 15).
Using the Hermes desktop app? See where plugins live

Full setup & settings ↗Source code ↗
More in memory & notes
New Hermes Agent updates every day on X. Follow @HermesWatcher